Broadcom Inc. (AVGO) is picking up serious momentum in the AI chip wars, and it has everything to do with Alphabet Inc. (GOOGL) deciding to go all-in on its homegrown Tensor Processing Units. Not just for internal use anymore, either. Google trained its Gemini 3 model entirely on TPUs and is now preparing to rent them out to external customers, which changes the game entirely.

BofA Securities analyst Vivek Arya just bumped his price target on Broadcom from $400 to $460, reiterating a Buy rating. His reasoning? Google's TPU expansion is a massive tailwind for Broadcom, which serves as the key silicon design partner behind these chips. When Google scales up, Broadcom scales up with it.

Here's where the numbers get interesting. Arya estimates the current TPU average selling price sits around $5,000 to $6,000, with roughly two million units shipped in 2025. By 2026, he expects that price to jump to $12,000 to $15,000 per unit, with shipments climbing past three million. If demand from Google's internal operations and new external customers like Anthropic and Meta Platforms Inc. (META) really takes off, shipments could stretch to 3.6 million to 3.8 million units.

The Performance Battle

So how does Google's TPUv7 actually stack up against Nvidia Corp's (NVDA) shiny new Blackwell Ultra GPUs? According to Arya's analysis, TPUv7 outperforms Nvidia's GB300 on power efficiency for specific AI training tasks, particularly FP8 large language models. The TPU delivers about 5.4 TFLOPs per watt versus roughly 3.6 TFLOPs per watt for Blackwell Ultra. That advantage translates to up to 40% lower total cost of ownership for optimized workloads.

But before anyone writes Nvidia's obituary, there's context. Performance varies dramatically depending on workload type, model support, and software optimization. And in those areas, Nvidia's GPU ecosystem still dominates. The company has spent years building out developer tools, frameworks, and partnerships that make its chips the default choice for most AI applications.

What's Coming Next?

Looking further out, the competitive landscape gets murkier. Nvidia's next major architecture, Vera Rubin, is expected in late 2026 and could leapfrog TPU on memory technology and system cost. Meanwhile, Broadcom's TPU roadmap may not deliver a major semiconductor upgrade until TPUv8 arrives in 2027, potentially narrowing those competitive gains.

Arya also flagged some real risks worth watching. If Google decides to shift from renting TPUs exclusively through Google Cloud to selling them directly to customers, it could create fresh competition for Broadcom's other custom Application-Specific Integrated Circuit customers, including Meta, ByteDance, and OpenAI. Google could also broaden its design partner roster by bringing in someone like MediaTek for lower-tier TPU variants, which would eat into Broadcom's share.

The Bigger Picture

Despite those uncertainties, Arya believes Broadcom remains well-positioned as AI compute spending continues to scale. He expects the company's custom ASIC and TPU products to capture up to 15% of the roughly $900 billion AI accelerator market by 2030, up from about 8% this year. That's not a small opportunity.

The analyst modeled more than $23 in Broadcom earnings per share by 2030 and modestly adjusted gross margin expectations to reflect a higher mix of compute silicon in the revenue base. Overall, he sees rising TPU leverage and broadening adoption as meaningful growth catalysts for Broadcom as AI infrastructure investment ramps through 2027 and beyond.

For the fourth quarter, Arya projected revenue of $17.445 billion, up slightly from his prior forecast of $17.405 billion. He adjusted his earnings estimate to $1.83 per share, down a couple cents from earlier guidance of $1.85.

Price Action: Broadcom stock was trading lower by 2.92% to $391.19 at publication on Monday.